@geravant/sinain 1.0.19 → 1.1.0

package/sinain-core/src/escalation/escalator.ts

@@ -0,0 +1,812 @@
+import type { AgentEntry, ContextWindow, EscalationConfig, OpenClawConfig, FeedItem, SpawnTaskMessage, SpawnTaskStatus } from "../types.js";
+import type { FeedBuffer } from "../buffers/feed-buffer.js";
+import type { WsHandler } from "../overlay/ws-handler.js";
+import type { Profiler } from "../profiler.js";
+import type { FeedbackStore } from "../learning/feedback-store.js";
+import type { SignalCollector } from "../learning/signal-collector.js";
+import { randomUUID, createHash } from "node:crypto";
+import { OpenClawWsClient } from "./openclaw-ws.js";
+import { EscalationSlot } from "./escalation-slot.js";
+import type { SlotEntry, QueueFeedbackCtx } from "./escalation-slot.js";
+import { shouldEscalate, calculateEscalationScore } from "./scorer.js";
+import { isCodingContext, buildEscalationMessage } from "./message-builder.js";
+import { loadPendingTasks, savePendingTasks, type PendingTaskEntry } from "../util/task-store.js";
+import { log, warn, error } from "../log.js";
+
+export interface HttpPendingEscalation {
+  id: string;
+  message: string;
+  score: number;
+  codingContext: boolean;
+  ts: number;
+  feedbackCtx: QueueFeedbackCtx | undefined;
+}
+
+const TAG = "escalation";
+
+export interface EscalatorDeps {
+  feedBuffer: FeedBuffer;
+  wsHandler: WsHandler;
+  escalationConfig: EscalationConfig;
+  openclawConfig: OpenClawConfig;
+  profiler?: Profiler;
+  feedbackStore?: FeedbackStore;
+  signalCollector?: SignalCollector;
+}
+
+/**
+ * Orchestrates escalation decisions and message delivery.
+ * Combines scorer (should we escalate?) + message builder (what to send) +
+ * OpenClaw WS delivery (how to send) into a single coordinator.
+ *
+ * Delivery uses a two-phase protocol:
+ *   Phase 1 (10s): await "accepted" frame → delivery confirmed, worker unblocks
+ *   Phase 2 (120s): await final frame → response arrives async, never trips circuit
+ *
+ * OutboundQueue persists messages to disk for crash recovery and
+ * provides content-hash idempotency keys for gateway-level dedup.
+ */
+export class Escalator {
+  private wsClient: OpenClawWsClient;
+  private slot: EscalationSlot;
+  private httpPending: HttpPendingEscalation | null = null;
+
+  private lastEscalationTs = Date.now();
+  private lastEscalatedDigest = "";
+
+  // Spawn deduplication state
+  private lastSpawnFingerprint = "";
+  private lastSpawnTs = 0;
+  private static readonly SPAWN_COOLDOWN_MS = 60_000; // 60 seconds between duplicate spawns
+
+  // Prevent concurrent spawn RPCs (sibling spawns only — never blocks regular escalations)
+  private spawnInFlight = false;
+
+  // Track pending spawn tasks for result fetching (persisted to disk)
+  private pendingSpawnTasks: Map<string, PendingTaskEntry>;
+
+  // Cap concurrent polling loops to limit RPC load
+  private static readonly MAX_CONCURRENT_POLLS = 5;
+  private activePolls = 0;
+  private pollQueue: string[] = [];
+
+  // Store context from last escalation for response handling
+  private lastEscalationContext: ContextWindow | null = null;
+
+  private stats = {
+    totalEscalations: 0,
+    totalResponses: 0,
+    totalErrors: 0,
+    totalNoReply: 0,
+    lastEscalationTs: 0,
+    lastResponseTs: 0,
+    // Health metrics
+    totalTimeouts: 0,
+    totalDirectResponses: 0,
+    totalSpawnResponses: 0,
+    avgResponseMs: 0,
+    consecutiveTimeouts: 0,
+    lastTimeoutTs: 0,
+  };
+
+  private outboundBytes = 0;
+
+  constructor(private deps: EscalatorDeps) {
+    this.wsClient = new OpenClawWsClient(deps.openclawConfig);
+    this.slot = new EscalationSlot(this.wsClient, deps.openclawConfig, {
+      onResponse: (result, entry, latencyMs) => this.handleEscalationResponse(result, entry, latencyMs),
+      onPhase1Failure: (isTimeout) => {
+        if (isTimeout) {
+          this.stats.totalTimeouts++;
+          this.stats.consecutiveTimeouts++;
+          this.stats.lastTimeoutTs = Date.now();
+          this.deps.profiler?.gauge("escalation.totalTimeouts", this.stats.totalTimeouts);
+          if (this.stats.consecutiveTimeouts >= 3) {
+            warn(TAG, `⚠ ${this.stats.consecutiveTimeouts} consecutive Phase 1 timeouts`);
+          }
+        }
+      },
+      onOutboundBytes: (n) => {
+        this.outboundBytes += n;
+        this.deps.profiler?.gauge("network.escalationOutBytes", this.outboundBytes);
+      },
+    });
+    // Load pending tasks from disk (crash recovery)
+    this.pendingSpawnTasks = loadPendingTasks();
+    // Attempt delivery on every WS reconnect
+    this.wsClient.on("connected", () => this.slot.onConnected());
+  }
+
+  /** Late-bind the signal collector (created after AgentLoop). */
+  setSignalCollector(sc: SignalCollector): void {
+    this.deps.signalCollector = sc;
+  }
+
+  /** Start the WS connection to OpenClaw (skipped when transport=http). */
+  start(): void {
+    if (this.deps.escalationConfig.mode !== "off" && this.deps.escalationConfig.transport !== "http") {
+      this.wsClient.connect();
+      const tokenHash = this.deps.openclawConfig.gatewayToken
+        ? createHash("sha256").update(this.deps.openclawConfig.gatewayToken).digest("hex").slice(0, 12)
+        : "none";
+      log(TAG, `mode: ${this.deps.escalationConfig.mode}, tokenHash: ${tokenHash}, wsUrl: ${this.deps.openclawConfig.gatewayWsUrl}`);
+    }
+  }
+
+  /** Stop and disconnect. */
+  stop(): void {
+    this.wsClient.disconnect();
+  }
+
+  /** Update escalation mode at runtime. */
+  setMode(mode: EscalatorDeps["escalationConfig"]["mode"]): void {
+    const wasOff = this.deps.escalationConfig.mode === "off";
+    this.deps.escalationConfig.mode = mode;
+    if (mode !== "off" && !this.wsClient.isConnected) {
+      this.wsClient.resetConnection();
+    }
+    if (mode === "off") {
+      this.wsClient.disconnect();
+    }
+    // Reset stale timer when transitioning from "off" to active (prevents immediate stale)
+    if (wasOff && mode !== "off") {
+      this.lastEscalationTs = Date.now();
+    }
+    log(TAG, `mode changed to: ${mode}`);
+  }
+
+  /**
+   * Called after every agent analysis tick.
+   * Decides whether to escalate and enqueues the message for delivery.
+   */
+  onAgentAnalysis(entry: AgentEntry, contextWindow: ContextWindow): void {
+    // Skip WS escalations when circuit is open (HTTP transport bypasses this)
+    const transport = this.deps.escalationConfig.transport;
+    if (this.wsClient.isCircuitOpen && transport !== "http") {
+      log(TAG, `tick #${entry.id}: skipped — circuit breaker open`);
+      return;
+    }
+
+    const { escalate, score, stale } = shouldEscalate(
+      entry.digest,
+      entry.hud,
+      contextWindow,
+      this.deps.escalationConfig.mode,
+      this.lastEscalationTs,
+      this.deps.escalationConfig.cooldownMs,
+      this.lastEscalatedDigest,
+      this.deps.escalationConfig.staleMs,
+    );
+
+    if (!escalate) {
+      log(TAG, `tick #${entry.id}: not escalating (mode=${this.deps.escalationConfig.mode}, score=${score.total}, hud="${entry.hud.slice(0, 40)}")`);
+      return;
+    }
+
+    // Mark cooldown immediately
+    this.stats.totalEscalations++;
+    this.deps.profiler?.gauge("escalation.totalEscalations", this.stats.totalEscalations);
+    this.lastEscalationTs = Date.now();
+    this.stats.lastEscalationTs = Date.now();
+    this.lastEscalatedDigest = entry.digest;
+
+    const staleTag = stale ? ", STALE" : "";
+    const wsState = this.wsClient.isConnected ? "ws=connected" : "ws=disconnected";
+    log(TAG, `escalating tick #${entry.id} (score=${score.total}, reasons=[${score.reasons.join(",")}]${staleTag}, ${wsState})`);
+
+    // Store context for response handling (used in pushResponse for coding-context max-length)
+    this.lastEscalationContext = contextWindow;
+
+    const escalationReason = score.reasons.join(", ");
+    const message = buildEscalationMessage(
+      entry.digest,
+      contextWindow,
+      entry,
+      this.deps.escalationConfig.mode,
+      escalationReason,
+    );
+
+    const slotId = createHash("sha256").update(this.deps.openclawConfig.sessionKey + entry.ts).digest("hex").slice(0, 16);
+    const slotEntry: SlotEntry = {
+      id: slotId,
+      message,
+      sessionKey: this.deps.openclawConfig.sessionKey,
+      feedbackCtx: {
+        tickId: entry.id,
+        hud: entry.hud,
+        currentApp: contextWindow.currentApp,
+        escalationScore: score.total,
+        escalationReasons: score.reasons,
+        codingContext: isCodingContext(contextWindow).coding,
+        digest: entry.digest,
+      },
+      ts: entry.ts,
+    };
+
+    const useHttp = transport === "http" || (transport === "auto" && !this.wsClient.isConnected);
+
+    if (useHttp) {
+      // Store in HTTP pending slot (newest wins, like EscalationSlot)
+      this.httpPending = {
+        id: slotId,
+        message,
+        score: score.total,
+        codingContext: isCodingContext(contextWindow).coding,
+        ts: entry.ts,
+        feedbackCtx: slotEntry.feedbackCtx,
+      };
+      log(TAG, `tick #${entry.id} → httpPending id=${slotId} (transport=${transport})`);
+    } else {
+      log(TAG, `tick #${entry.id} → slot.insert id=${slotId} depth=${this.slot.depth}`);
+      this.slot.insert(slotEntry);
+    }
+  }
+
+  /** Push fresh SITUATION.md content to the gateway server (fire-and-forget). */
+  pushSituationMd(content: string): void {
+    if (!this.wsClient.isConnected) return;
+    this.wsClient.sendRpc("situation.update", { content }, 10_000)
+      .catch((err: any) => warn(TAG, `situation.update rpc failed: ${err.message}`));
+  }
+
+  /** Send a direct user message to OpenClaw. */
+  async sendDirect(text: string): Promise<void> {
+    const idemKey = `direct-${Date.now()}`;
+    if (this.wsClient.isConnected) {
+      try {
+        await this.wsClient.sendAgentRpc(text, idemKey, this.deps.openclawConfig.sessionKey);
+        return;
+      } catch (err: any) {
+        warn(TAG, `sendDirect RPC failed: ${err.message}`);
+      }
+    }
+    // WS disconnected or RPC failed — surface error to HUD
+    const errMsg = `[⚠] Gateway disconnected — message not sent`;
+    this.deps.feedBuffer.push(errMsg, "normal", "openclaw", "stream");
+    this.deps.wsHandler.broadcast(errMsg, "normal", "stream");
+  }
+
+  /**
+   * Send a periodic feedback summary to the OpenClaw agent.
+   * Called on a timer from index.ts when learning is enabled.
+   * Returns true if the summary was sent successfully.
+   */
+  async sendFeedbackSummary(): Promise<boolean> {
+    if (!this.deps.feedbackStore) return false;
+    if (!this.wsClient.isConnected) return false;
+
+    const stats = this.deps.feedbackStore.getStats();
+    const totalRecords = stats.totalRecords as number;
+    if (totalRecords < 3) return false;
+
+    const recent = this.deps.feedbackStore.queryRecent(5);
+    const withSignals = recent.filter(r => r.signals.compositeScore !== 0 || r.signals.errorCleared !== null);
+    if (withSignals.length === 0) return false;
+
+    // Format compact summary
+    const topTags = (stats.topTags as [string, number][] || [])
+      .slice(0, 5)
+      .map(([tag, count]) => `${tag} (${count})`)
+      .join(", ");
+
+    const recentLines = withSignals.slice(0, 5).map(r => {
+      const ok = r.signals.compositeScore >= 0.2;
+      const icon = ok ? "✓" : "✗";
+      const score = r.signals.compositeScore.toFixed(2);
+      const tags = r.tags.slice(0, 3).join(", ");
+      const details: string[] = [];
+      if (r.signals.errorCleared === true) details.push("error cleared");
+      if (r.signals.errorCleared === false) details.push("error persisted");
+      if (r.signals.noReEscalation === true) details.push("no re-escalation");
+      if (r.signals.noReEscalation === false) details.push("re-escalated");
+      if (r.signals.quickAppSwitch === true) details.push("quick switch");
+      return `  ${icon} ${score} [${tags}]${details.length > 0 ? " — " + details.join(", ") : ""}`;
+    });
+
+    const message = `[sinain-core:feedback-summary]
+
+Escalations: ${totalRecords} | Avg score: ${stats.avgCompositeScore ?? "n/a"} | Avg latency: ${stats.avgLatencyMs ?? "n/a"}ms
+Top tags: ${topTags || "none"}
+
+Recent (last ${withSignals.length}):
+${recentLines.join("\n")}`;
+
+    const idemKey = `feedback-summary-${Date.now()}`;
+    try {
+      await this.wsClient.sendAgentRpc(message, idemKey, this.deps.openclawConfig.sessionKey);
+      log(TAG, `feedback summary sent (${totalRecords} records, ${withSignals.length} with signals)`);
+      return true;
+    } catch (err: any) {
+      warn(TAG, `feedback summary send failed: ${err.message}`);
+      return false;
+    }
+  }
+
+  /** Return the current HTTP pending escalation (or null). */
+  getPendingHttp(): HttpPendingEscalation | null {
+    return this.httpPending;
+  }
+
+  /** Respond to an HTTP pending escalation. */
+  respondHttp(id: string, response: string): { ok: boolean; error?: string } {
+    if (!this.httpPending) {
+      return { ok: false, error: "no pending escalation" };
+    }
+    if (this.httpPending.id !== id) {
+      return { ok: false, error: `id mismatch: expected ${this.httpPending.id}` };
+    }
+
+    this.pushResponse(response, this.lastEscalationContext);
+
+    // Record feedback (async, non-blocking)
+    if (this.httpPending.feedbackCtx) {
+      const { digest, ...ctx } = this.httpPending.feedbackCtx;
+      this.recordFeedback(ctx, digest, this.httpPending.message, response, Date.now() - this.httpPending.ts);
+    }
+
+    log(TAG, `httpPending id=${id} responded (${response.length} chars)`);
+    this.httpPending = null;
+    return { ok: true };
+  }
+
+  /** Whether the gateway WS client is currently connected. */
+  get isGatewayConnected(): boolean {
+    return this.wsClient.isConnected;
+  }
+
+  /** Force-reconnect the gateway WS client. */
+  reconnectGateway(): void {
+    this.wsClient.resetConnection();
+  }
+
+  /** Get stats for /health. */
+  getStats(): Record<string, unknown> {
+    return {
+      mode: this.deps.escalationConfig.mode,
+      transport: this.deps.escalationConfig.transport,
+      gatewayConnected: this.wsClient.isConnected,
+      circuitOpen: this.wsClient.isCircuitOpen,
+      slotDepth: this.slot.depth,
+      slotInFlight: this.slot.inFlightId,
+      httpPendingId: this.httpPending?.id ?? null,
+      spawnInFlight: this.spawnInFlight,
+      cooldownMs: this.deps.escalationConfig.cooldownMs,
+      staleMs: this.deps.escalationConfig.staleMs,
+      pendingSpawnTasks: this.pendingSpawnTasks.size,
+      ...this.stats,
+    };
+  }
+
+  /**
+   * Dispatch a task to a spawned subagent via direct child session addressing.
+   * Creates a unique child session key and sends the task directly to the gateway
+   * agent RPC — bypassing the main session to avoid dedup/NO_REPLY issues.
+   */
+  async dispatchSpawnTask(task: string, label?: string): Promise<void> {
+    // Prevent sibling spawn RPCs from piling up (independent from escalation queue)
+    if (this.spawnInFlight) {
+      log(TAG, `spawn-task skipped — spawn RPC already in-flight`);
+      return;
+    }
+
+    // --- Fingerprint dedup — hash the task content ---
+    const fingerprint = createHash("sha256").update(task.trim()).digest("hex").slice(0, 16);
+    const now = Date.now();
+
+    if (fingerprint === this.lastSpawnFingerprint &&
+        now - this.lastSpawnTs < Escalator.SPAWN_COOLDOWN_MS) {
+      log(TAG, `spawn-task skipped (duplicate fingerprint ${fingerprint})`);
+      return;
+    }
+
+    this.lastSpawnFingerprint = fingerprint;
+    this.lastSpawnTs = now;
+
+    const taskId = `spawn-${Date.now()}`;
+    const startedAt = Date.now();
+    const labelStr = label ? ` (label: "${label}")` : "";
+    const idemKey = `spawn-task-${Date.now()}`;
+
+    // Generate a unique child session key — bypasses the main agent entirely
+    const childSessionKey = `agent:main:subagent:${randomUUID()}`;
+    const mainSessionKey = this.deps.openclawConfig.sessionKey;
+
+    this.outboundBytes += Buffer.byteLength(task);
+    this.deps.profiler?.gauge("network.escalationOutBytes", this.outboundBytes);
+    log(TAG, `dispatching spawn-task${labelStr} → child=${childSessionKey}: "${task.slice(0, 80)}..."`);
+
+    // ★ Broadcast "spawned" BEFORE the RPC — TSK tab shows ··· immediately
+    this.broadcastTaskEvent(taskId, "spawned", label, startedAt);
+
+    if (!this.wsClient.isConnected) {
+      warn(TAG, `spawn-task ${taskId}: WS disconnected — cannot dispatch`);
+      this.broadcastTaskEvent(taskId, "failed", label, startedAt);
+      return;
+    }
+
+    // ★ Set spawnInFlight BEFORE first await — cleared in finally regardless of outcome.
+    // Dedicated lane flag: never touches the escalation queue so regular escalations
+    // continue unblocked while this spawn RPC is pending.
+    this.spawnInFlight = true;
+    try {
+      // Send directly to a new child session via the gateway agent RPC
+      const result = await this.wsClient.sendRpc("agent", {
+        message: task,
+        sessionKey: childSessionKey,
+        lane: "subagent",
+        extraSystemPrompt: this.buildChildSystemPrompt(task, label),
+        deliver: false,
+        spawnedBy: mainSessionKey,
+        idempotencyKey: idemKey,
+        label: label || undefined,
+      }, 45_000, { expectFinal: true });
+
+      log(TAG, `spawn-task RPC response: ${JSON.stringify(result).slice(0, 500)}`);
+      this.stats.totalSpawnResponses++;
+
+      // Extract result — child agent actually ran the task and returned content
+      const payloads = result?.payload?.result?.payloads;
+      const runId = result?.payload?.runId || taskId;
+
+      if (Array.isArray(payloads) && payloads.length > 0) {
+        const output = payloads.map((pl: any) => pl.text || "").join("\n").trim();
+        if (output) {
+          this.pushResponse(`${label || "Background task"}:\n${output}`);
+          this.broadcastTaskEvent(taskId, "completed", label, startedAt, output);
+        } else {
+          log(TAG, `spawn-task: ${payloads.length} payloads but empty text, trying chat.history`);
+          const historyText = await this.fetchChildResult(childSessionKey);
+          this.broadcastTaskEvent(taskId, "completed", label, startedAt,
+            historyText || "task completed (no output)");
+          if (historyText) {
+            this.pushResponse(`${label || "Background task"}:\n${historyText}`);
+          }
+        }
+      } else {
+        // No payloads — fallback: fetch from chat.history on child session
+        log(TAG, `spawn-task: no payloads, fetching chat.history for child=${childSessionKey}`);
+        const historyText = await this.fetchChildResult(childSessionKey);
+        if (historyText) {
+          this.pushResponse(`${label || "Background task"}:\n${historyText}`);
+          this.broadcastTaskEvent(taskId, "completed", label, startedAt, historyText);
+        } else {
+          this.broadcastTaskEvent(taskId, "completed", label, startedAt,
+            "task completed (no output captured)");
+        }
+      }
+
+      // Persist for crash recovery (no polling needed — result already in hand)
+      this.pendingSpawnTasks.set(taskId, {
+        runId,
+        childSessionKey,
+        label,
+        startedAt,
+        pollingEmitted: false,
+      });
+      savePendingTasks(this.pendingSpawnTasks);
+
+      // Clean up immediately since we already have the result
+      this.pendingSpawnTasks.delete(taskId);
+      savePendingTasks(this.pendingSpawnTasks);
+    } catch (err: any) {
+      error(TAG, `spawn-task failed: ${err.message}`);
+      this.broadcastTaskEvent(taskId, "failed", label, startedAt);
+    } finally {
+      this.spawnInFlight = false;
+    }
+  }
+
+  /** Build a focused system prompt for the child subagent. */
+  private buildChildSystemPrompt(task: string, label?: string): string {
+    return [
+      "# Subagent Context",
+      "",
+      "You are a **subagent** spawned for a specific task.",
+      "",
+      "## Your Role",
+      `- Task: ${task.replace(/\s+/g, " ").trim().slice(0, 500)}`,
+      "- Complete this task. That's your entire purpose.",
+      "",
+      "## Rules",
+      "1. Stay focused — do your assigned task, nothing else",
+      "2. Your final message will be reported to the requester",
+      "3. Be concise but informative",
+      "",
+      label ? `Label: ${label}` : "",
+    ].filter(Boolean).join("\n");
+  }
+
+  /** Fetch the latest assistant reply from a child session's chat history. */
+  private async fetchChildResult(childSessionKey: string): Promise<string | null> {
+    try {
+      const historyResult = await this.wsClient.sendRpc("chat.history", {
+        sessionKey: childSessionKey,
+        limit: 10,
+      }, 10_000);
+      return this.extractLatestAssistantReply(historyResult);
+    } catch (err: any) {
+      warn(TAG, `chat.history fetch failed for ${childSessionKey}: ${err.message}`);
+      return null;
+    }
+  }
+
+  /** Poll for task completion and push result to HUD (preserved for crash recovery). */
+  private async pollTaskCompletion(taskId: string): Promise<void> {
+    // Enforce concurrency cap — queue excess tasks
+    if (this.activePolls >= Escalator.MAX_CONCURRENT_POLLS) {
+      log(TAG, `poll queued (${this.activePolls} active): taskId=${taskId}`);
+      this.pollQueue.push(taskId);
+      return;
+    }
+
+    this.activePolls++;
+    this.deps.profiler?.gauge("escalation.activePolls", this.activePolls);
+
+    const task = this.pendingSpawnTasks.get(taskId);
+    if (!task) {
+      this.finishPoll();
+      return;
+    }
+
+    const maxWaitMs = 5 * 60 * 1000; // 5 minutes
+    const pollIntervalMs = 5000; // 5 seconds
+
+    const poll = async (): Promise<void> => {
+      const elapsed = Date.now() - task.startedAt;
+      if (elapsed > maxWaitMs) {
+        log(TAG, `spawn-task timeout: taskId=${taskId}`);
+        this.broadcastTaskEvent(taskId, "timeout", task.label, task.startedAt);
+        this.pendingSpawnTasks.delete(taskId);
+        savePendingTasks(this.pendingSpawnTasks);
+        this.finishPoll();
+        return;
+      }
+
+      if (!this.wsClient.isConnected) {
+        // Retry later
+        setTimeout(() => poll(), pollIntervalMs);
+        return;
+      }
+
+      try {
+        // Wait for completion (short timeout to poll periodically)
+        const waitResult = await this.wsClient.sendRpc("agent.wait", {
+          runId: task.runId,
+          timeoutMs: pollIntervalMs,
+        }, pollIntervalMs + 2000);
+
+        // Debug: log the poll result
+        log(TAG, `poll result: taskId=${taskId}, status=${waitResult?.payload?.status}, ok=${waitResult?.ok}`);
+
+        // Accept multiple completion statuses
+        const completedStatuses = ["ok", "completed", "done", "finished", "success"];
+        const status = waitResult?.payload?.status;
+
+        if (waitResult?.ok && completedStatuses.includes(status)) {
+          log(TAG, `spawn-task completed: taskId=${taskId}, status=${status}`);
+
+          // Fetch the result from chat history
+          const historyResult = await this.wsClient.sendRpc("chat.history", {
+            sessionKey: task.childSessionKey,
+            limit: 10,
+          }, 10000);
+
+          const resultText = this.extractLatestAssistantReply(historyResult);
+          if (resultText) {
+            const labelDisplay = task.label || "Background task";
+            this.pushResponse(`${labelDisplay}:\n${resultText}`);
+          } else {
+            log(TAG, `spawn-task completed but no result text: taskId=${taskId}`);
+          }
+
+          this.broadcastTaskEvent(taskId, "completed", task.label, task.startedAt, resultText ?? undefined);
+          this.pendingSpawnTasks.delete(taskId);
+          savePendingTasks(this.pendingSpawnTasks);
+          this.finishPoll();
+          return;
+        }
+
+        if (waitResult?.payload?.status === "error" || waitResult?.payload?.status === "failed") {
+          log(TAG, `spawn-task failed: taskId=${taskId}, error=${waitResult?.payload?.error || "unknown"}`);
+          this.broadcastTaskEvent(taskId, "failed", task.label, task.startedAt);
+          this.pendingSpawnTasks.delete(taskId);
+          savePendingTasks(this.pendingSpawnTasks);
+          this.finishPoll();
+          return;
+        }
+
+        // Status is "timeout" or still running — emit polling once
+        if (!task.pollingEmitted) {
+          task.pollingEmitted = true;
+          this.broadcastTaskEvent(taskId, "polling", task.label, task.startedAt);
+        }
+        setTimeout(() => poll(), 1000);
+      } catch (err: any) {
+        warn(TAG, `poll error for taskId=${taskId}: ${err.message}`);
+        // Retry on transient errors
+        setTimeout(() => poll(), pollIntervalMs);
+      }
+    };
+
+    // Start polling
+    poll();
+  }
+
+  /** Decrement active polls and drain the queue. */
+  private finishPoll(): void {
+    this.activePolls--;
+    this.deps.profiler?.gauge("escalation.activePolls", this.activePolls);
+    this.deps.profiler?.gauge("escalation.pendingSpawns", this.pendingSpawnTasks.size);
+    // Drain queued tasks
+    while (this.pollQueue.length > 0 && this.activePolls < Escalator.MAX_CONCURRENT_POLLS) {
+      const nextId = this.pollQueue.shift()!;
+      if (this.pendingSpawnTasks.has(nextId)) {
+        log(TAG, `poll dequeued: taskId=${nextId}`);
+        this.pollTaskCompletion(nextId);
+      }
+    }
+  }
+
+  /** Extract the latest assistant reply from chat history. */
+  private extractLatestAssistantReply(historyResult: any): string | null {
+    // Try multiple paths to find messages (different API response formats)
+    const messages = historyResult?.payload?.messages
+      || historyResult?.messages
+      || historyResult?.payload?.result?.messages
+      || historyResult?.result?.messages;
+
+    // Debug: log what we found
+    log(TAG, `extractLatestAssistantReply: messages=${Array.isArray(messages) ? messages.length : "none"}`);
+
+    if (!Array.isArray(messages)) {
+      // Maybe it's a direct text response
+      if (typeof historyResult?.payload?.text === "string") {
+        log(TAG, `extractLatestAssistantReply: found payload.text`);
+        return historyResult.payload.text;
+      }
+      if (typeof historyResult?.text === "string") {
+        log(TAG, `extractLatestAssistantReply: found text`);
+        return historyResult.text;
+      }
+      if (typeof historyResult?.payload?.result?.text === "string") {
+        log(TAG, `extractLatestAssistantReply: found payload.result.text`);
+        return historyResult.payload.result.text;
+      }
+      log(TAG, `extractLatestAssistantReply: no messages array found, historyResult keys=${Object.keys(historyResult || {}).join(",")}`);
+      return null;
+    }
+
+    // Find the last assistant message
+    for (let i = messages.length - 1; i >= 0; i--) {
+      const msg = messages[i];
+      if (msg?.role === "assistant") {
+        // Extract text content
+        if (typeof msg.content === "string") return msg.content;
+        if (Array.isArray(msg.content)) {
+          const textPart = msg.content.find((p: any) => p.type === "text");
+          if (textPart?.text) return textPart.text;
+        }
+      }
+    }
+
+    log(TAG, `extractLatestAssistantReply: no assistant message found in ${messages.length} messages`);
+    return null;
+  }
+
+  // ── Private ──
+
+  /** Process the agent response arriving in Phase 2 (called by EscalationSlot callback). */
+  private handleEscalationResponse(result: any, entry: SlotEntry, rpcLatencyMs: number): void {
+    if (result?.ok && result.payload) {
+      const p = result.payload;
+      log(TAG, `WS RPC ok → runId=${p.runId}, status=${p.status}, latency=${rpcLatencyMs}ms`);
+
+      this.stats.totalDirectResponses++;
+      this.stats.consecutiveTimeouts = 0;
+      // EMA α=0.2: smooths latency while reacting to sustained changes
+      this.stats.avgResponseMs = this.stats.avgResponseMs === 0
+        ? rpcLatencyMs
+        : this.stats.avgResponseMs * 0.8 + rpcLatencyMs * 0.2;
+
+      const payloads = p.result?.payloads;
+      let responseText = "";
+      if (Array.isArray(payloads) && payloads.length > 0) {
+        const output = payloads.map((pl: any) => pl.text || "").join("\n").trim();
+        responseText = output;
+        if (output && !output.startsWith("NO_REPLY")) {
+          this.pushResponse(output, this.lastEscalationContext);
+        } else {
+          this.stats.totalNoReply++;
+          this.deps.profiler?.gauge("escalation.totalNoReply", this.stats.totalNoReply);
+          log(TAG, output ? `agent returned NO_REPLY as text — silent` : `empty text in ${payloads.length} payloads`);
+        }
+      } else {
+        this.stats.totalNoReply++;
+        this.deps.profiler?.gauge("escalation.totalNoReply", this.stats.totalNoReply);
+        log(TAG, "agent returned NO_REPLY — silent");
+      }
+
+      // Record feedback (async, non-blocking)
+      if (entry.feedbackCtx) {
+        const { digest, ...ctx } = entry.feedbackCtx;
+        this.recordFeedback(ctx, digest, entry.message, responseText, rpcLatencyMs);
+      }
+    } else if (result && !result.ok) {
+      const errDetail = JSON.stringify(result.error || result.payload);
+      log(TAG, `agent RPC error: ${errDetail}`);
+      this.pushError(errDetail);
+      this.stats.totalErrors++;
+      this.deps.profiler?.gauge("escalation.errors", this.stats.totalErrors);
+    }
+  }
+
+  private broadcastTaskEvent(
+    taskId: string,
+    status: SpawnTaskStatus,
+    label?: string,
+    startedAt?: number,
+    resultPreview?: string,
+  ): void {
+    const now = Date.now();
+    const isTerminal = status === "completed" || status === "failed" || status === "timeout";
+    const msg: SpawnTaskMessage = {
+      type: "spawn_task",
+      taskId,
+      label: label || "Background task",
+      status,
+      startedAt: startedAt || now,
+      ...(isTerminal ? { completedAt: now } : {}),
+      ...(resultPreview ? { resultPreview: resultPreview.slice(0, 200) } : {}),
+    };
+    log(TAG, `broadcast spawn_task: taskId=${taskId}, status=${status}, clients=${this.deps.wsHandler.clientCount}`);
+    this.deps.wsHandler.broadcastRaw(msg);
+  }
+
+  private pushResponse(output: string, context?: ContextWindow | null): void {
+    // Allow longer responses for coding contexts
+    const { coding } = context ? isCodingContext(context) : { coding: false };
+    const maxLen = coding ? 4000 : 3000;
+
+    const text = `[🤖] ${output.trim().slice(0, maxLen)}`;
+    this.deps.feedBuffer.push(text, "high", "openclaw", "agent");
+    this.deps.wsHandler.broadcast(text, "high", "agent");
+    this.stats.totalResponses++;
+    this.deps.profiler?.gauge("escalation.totalResponses", this.stats.totalResponses);
+    this.stats.lastResponseTs = Date.now();
+    log(TAG, `response pushed (coding=${coding}, maxLen=${maxLen}): "${output.slice(0, 80)}..."`);
+  }
+
+  private pushError(detail: string): void {
+    const text = `[\ud83e\udd16 err] ${detail.slice(0, 500)}`;
+    this.deps.feedBuffer.push(text, "normal", "openclaw", "stream");
+  }
+
+  /** Record a feedback entry after successful escalation. Safe — never throws. */
+  private recordFeedback(
+    ctx: { tickId: number; hud: string; currentApp: string; escalationScore: number; escalationReasons: string[]; codingContext: boolean } | undefined,
+    digest: string,
+    escalationMessage: string,
+    openclawResponse: string,
+    responseLatencyMs: number,
+  ): void {
+    if (!ctx || !this.deps.feedbackStore || !this.deps.signalCollector) return;
+    try {
+      const record = this.deps.feedbackStore.createRecord({
+        tickId: ctx.tickId,
+        digest,
+        hud: ctx.hud,
+        currentApp: ctx.currentApp,
+        escalationScore: ctx.escalationScore,
+        escalationReasons: ctx.escalationReasons,
+        codingContext: ctx.codingContext,
+        escalationMessage,
+        openclawResponse,
+        responseLatencyMs,
+      });
+      this.deps.feedbackStore.append(record);
+      this.deps.signalCollector.schedule(record);
+    } catch (err: any) {
+      warn(TAG, `feedback record failed: ${err.message}`);
+    }
+  }
+}